U.S. patent number 5,668,888 [Application Number 08/174,175] was granted by the patent office on 1997-09-16 for method and system for automatic detection of ribs and pneumothorax in digital chest radiographs.
This patent grant is currently assigned to Arch Development Corporation. Invention is credited to Kunio Doi, Shigeru Sanada.
United States Patent |
5,668,888 |
Doi , et al. |
September 16, 1997 |
Method and system for automatic detection of ribs and pneumothorax
in digital chest radiographs
Abstract
A method and system for the detection of anatomical features in
a digital chest radiograph, and in particular rib edges and
pneumothorax, wherein vertical profiles are defined in regions of
the lungs in the radiograph, and a model function is fitted to the
profiles to obtain initial estimates of the locations of rib edges.
Gradient-histogram analysis is performed by determining edge
gradient and their corresponding orientations in regions of
interest (ROIs) defined in the radiograph in the vicinity of the
initially estimated rib edges, determining the maximum edge
gradient in each of the ROIs, and forming histograms which define
accurate detections of rib edges. The edges are fitted with an
elliptical function and a representation of the entire rib cage
structure is obtained. The method and system are also applied to
the detection of pneumothorax where ROIs are defined in the apical
lung region and edge gradients and their orientations are detected
in the ROIs. The ROI is edge-enhanced using the edge gradients and
their orientations and previously detected rib edges are removed.
Subtle, curved-line structure corresponding to pneumothorax in the
edge-enhanced ROI are detected and a representation of the
pneumothorax is obtained using a Hough transform.
Inventors: |
Doi; Kunio (Willowbrook,
IL), Sanada; Shigeru (Iwade-machi, JP) |
Assignee: |
Arch Development Corporation
(Chicago, IL)
|
Family
ID: |
24472165 |
Appl.
No.: |
08/174,175 |
Filed: |
December 29, 1993 |
Related U.S. Patent Documents
|
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
Issue Date |
|
|
617080 |
Nov 21, 1990 |
|
|
|
|
Current U.S.
Class: |
382/132;
382/199 |
Current CPC
Class: |
G06T
5/50 (20130101); G06T 7/12 (20170101); G06T
7/149 (20170101); G06T 7/174 (20170101); G06T
2207/10116 (20130101); G06T 2207/20116 (20130101); G06T
2207/20224 (20130101); G06T 2207/30008 (20130101); G06T
2207/30061 (20130101) |
Current International
Class: |
G06T
5/00 (20060101); G06G 007/60 () |
Field of
Search: |
;364/413.13
;382/6,9,16,18,22,54,132,199 |
References Cited
[Referenced By]
U.S. Patent Documents
Other References
Proceedings of the SPIE--International Society for Optical Eng.,
vol. 1002, pp. 158-167, 1989, Gregson et al., "A
symmetry-insensitive edge enhancement filter . . . ", (abstract
only). .
Proceedings of the SPIE--International Society for Optical Eng.,
vol. 1660, pp. 446-454, 1992, Kogelmeyer et al., "Computer
Detection of Stellate Lessions in Mamograms" (abstract only). .
Computer Vision, Ballard et al., Prentice-Hall, Inc., New Jersey,
1982, pp. 63-148. .
Med. Phys., issue 19/5, 1992, Sanada et al., "Image Feature
Analysis and Computer-aided Diagnosis in Digital Radiography:
Automated Detection of Pneumothorax in Chest Images", pp. 1153-1160
(abstract only provided). .
Digital Image Processing, 2.sup.nd Edition, Addison-Wesley Pub.
Co., 1987, Gonzalez et al., pp. 331-390. .
Radiology, Apr. 1988, Goodman et al., "Pneumothorax and other Lung
Diseases Effect of Altered Resolution and Edge Enhancement on
Diagnosis with Digitized Radiographs", pp. 83-88. .
Proceedings of ISM III '82, CH1804-4/82/0000/0114, 1982, Cocklin et
al., "Digital Enhancement of Pneumothoraces" pp. 114-116 (abstract
only)..
|
Primary Examiner: Weinhardt; Robert A.
Assistant Examiner: Thomas; Joseph
Attorney, Agent or Firm: Oblon, Spivak, McClelland, Maier
& Neustadt, P.C.
Government Interests
The U.S. Government has a paid-up license in this invention and the
right in limited circumstances to require the patent owner to
license others on reasonable terms as provided for by the terms of
the contracts nos. USPHS CA 24806, USPHS CA 47043, and USPHS CA
41851 awarded by the National Institute of Health.
Parent Case Text
This application is a Continuation of application Ser. No.
07/617,080, filed on Nov. 21, 1990, now abandoned.
Claims
What is claimed as new and desired to be secured by Letters Patent
of the U.S. is:
1. A method for the detection of a anatomic feature in a chest
radiograph, comprising:
(a) defining plural regions of interest in said radiograph;
(b) determining edge gradients in at least one of said plural
regions of interest and corresponding orientations of said edge
gradients for each of a plurality of pixels included in said at
least one of said regions of interest;
(c) determining a cumulative edge gradient of the edge gradients
and corresponding orientations determined in said at least one of
said plural regions of interest in step (b) by summing in said at
least one of said plural regions of interest the edge gradients
having the same orientation; and
(d) detecting at least one edge of said anatomic feature based upon
the cumulative edge gradient determined in said at least one of
said plural regions of interest, comprising,
(d1) determining which regions of interest in aid cumulative edge
gradient have an orientation within a predetermined range of
orientations; and
(d2) determining a location of an edge of said anatomic feature
based on locations of regions of interest which are determined in
step (d1) to have said orientation within said predetermined range
of orientations; and
(e) producing a representation of said anatomic feature using said
at least one detected edge.
2. A method as recited in claim 1, wherein said step (b) of
determining said edge gradients comprises:
fitting said pixels in said regions of interest using a
predetermined function;
subtracting said predetermined function from said pixels to obtain
background-trend-corrected regions of interest; and
determining said edge gradients using pixels in said
background-trend-corrected regions of interest.
3. A system according to claim 1, wherein:
step (b) comprises determining edge gradients and corresponding
orientations of said edge gradients for each of the plurality of
pixels included in each of said plural regions of interest;
step (c) comprises determining a respective cumulative edge
gradient for each of said plural regions of interest; and
step (d1) comprises determining which of said cumulative edge
gradients has a largest sum within said predetermined range of
orientations; and
step (d2) comprises determining said edge of said anatomic feature
as a location of the region of interest having in its respective
cumulative edge gradient said largest sum of orientations within
said predetermined range of orientations.
4. A method as recited in claim 3, wherein rib edges are detected,
a vertical direction in said chest radiograph is a direction
generally parallel to a subject's spine and a horizontal direction
is a direction perpendicular to said vertical direction, and said
step (a) of defining regions of interest comprises:
determining vertical profiles in regions of both lungs of said
chest radiograph;
fitting a predetermined model function to said vertical profiles to
obtain initial estimates of said rib edges in said radiograph;
and
defining said regions of interest in said radiograph to include
regions containing said initial estimates.
5. A method as recited in claim 4, further comprising:
determining a magnitude of each of said determined edge gradients
in each of said regions of interest;
determining corresponding orientations of said magnitudes of said
edge gradients in a polar coordinate system having r and .theta. as
variables;
determining said cumulative edge gradients in each region of
interest by summing said magnitudes of said edge gradients having
the same edge gradient orientation in the respective region of
interest; and
forming histograms of said cumulative edge gradients and using said
histograms to define upper and lower edges of a rib.
6. A method as recited in claim 5, wherein said step of defining
said regions of interest comprises:
defining a plurality of said regions of interest in a vertical
orientation with respect to said radiograph and with respect to
each other; and
overlapping each of said plurality of regions of interest by no
more than half of an adjacent one of said plurality of regions of
interest.
7. A method as recited in claim 4, wherein said step of producing a
representation comprises:
fitting each of said detected rib edges with an elliptical
function; and
obtaining a representation of an entire rib structure using said
rib edges fitted with said elliptical function.
8. A method as recited in claim 7, wherein said step of fitting
each of said rib edges further comprises:
grouping detected rib edge points of a selected rib based upon
orientation of said edge gradients of said selected rib for
distinction of upper and lower edges of said selected rib, and upon
vertical positions of said detected rib edge points of said
selected rib as a function of horizontal position in said
radiograph.
9. A method as recited in claim 7, further comprising:
estimating a set of fitting parameters for a missing of said
detected rib edges to obtain an estimated rib edge based upon
previously fitted rib edges; and
reperforming said steps of defining regions of interest,
determining edge gradients and detecting rib edges.
10. A system for the detection of an anatomic feature in a chest
radiograph, comprising:
first means for defining regions of interest in said
radiograph;
second means for determining edge gradients in said regions of
interest and corresponding orientations of said edge gradients for
each of a plurality of pixels included in each of said regions of
interest;
third means for determining a respective cumulative edge gradient
for each region of interest by summing in each region of interest
the edge gradients having the same orientation;
fourth means for detecting at least one edge of said anatomic
feature based upon the cumulative edge gradients determined in said
plural regions of interest, comprising,
means for determining which of said cumulative edge gradients has a
largest sum within a predetermined range of orientations; and
means for determining an edge of said anatomic region based on a
location of the region of interest having in its respective
cumulative edge gradient said largest sum within said predetermined
range of orientations; and
fifth means for producing a representation of said anatomic feature
using said at least one detected edge.
11. A system as recited in claim 10, wherein said second means for
determining edge gradients comprises:
means for fitting said pixels in said regions of interest using a
predetermined function;
means for subtracting said predetermined function from said pixels
to obtain background-trend-corrected regions of interest; and
means for determining said edge gradients using pixels in said
background-trend-corrected regions of interest.
12. A system as recited in claim 10, wherein said anatomic feature
includes rib edges, a vertical direction in said chest radiograph
is a direction generally parallel to the subject's spine and a
horizontal direction is a direction perpendicular to said vertical
direction, and said first means comprises:
means for determining vertical profiles in both lung regions of
said chest radiograph;
means for fitting a predetermined model function to said vertical
profiles to obtain initial estimates of said rib edges in said
radiograph; and
means for defining said regions of interest in said radiograph to
include regions containing said initial estimates.
13. A system as recited in claim 12, wherein said fifth means
comprises:
means for fitting each of said detected rib edges with an
elliptical function; and
means for obtaining a representation of entire ribs using said rib
edges fitted with said elliptical function.
14. A system as recited in claim 12, wherein said third means
comprises:
means for determining a magnitude of each of said determined edge
gradients in each of said regions of interest;
means for plotting said magnitudes of said edge gradients and
corresponding orientations of said magnitudes of said edge
gradients in a polar coordinate system having r and .theta. as
variables, wherein said magnitudes of said edge gradients are
plotted as said r variable and said corresponding orientations of
said edge gradient magnitudes are plotted as said .theta.
variable;
summing means for summing in each region of interest the magnitudes
of said edge gradients having the same orientation to produce said
cumulative edge gradients; and
means for forming histograms of said cumulative edge gradients and
using said histograms to define upper and lower edges of a rib.
15. A system as recited in claim 14, wherein said first means for
defining said regions of interest comprises:
means for defining a plurality of said regions of interest in a
vertical orientation; and
means for overlapping each of said plurality of regions of interest
by no more than half of an adjacent of said plurality of regions of
interest.
16. A system as recited in claim 13, wherein said means for fitting
each of said rib edges further comprises:
means for grouping detected rib edge points of a selected rib based
upon orientation of said edge gradients of said selected rib for
distinction of upper and lower edges of said selected rib, and upon
vertical positions of said detected rib edge points of said
selected rib as a function of horizontal position in said
radiograph.
17. A system as recited in claim 13, further comprising:
means for estimating a set of fitting parameters for a missing rib
edge to obtain an estimated rib edge based upon previously fitted
rib edges.
18. A method for detection of a pneumothorax in a chest radiograph,
comprising:
detecting rib edges in said radiograph;
defining a pneumothorax region of interest in said radiograph which
includes said pneumothorax;
removing said rib edges and respective areas adjacent said rib
edges from said pneumothorax region of interest;
detecting a curved line in said pneumothorax region of interest
having said rib edges removed; and
producing a representation of said pneumothorax based on the
detected curved line;
wherein said step of detecting rib edges comprises:
defining plural regions of interest in said radiograph which
include regions containing rib edges;
determining edge gradients and corresponding orientations of the
determined edge gradients for each of a plurality of pixels
included in each of said regions of interest;
determining a respective cumulative edge gradient for each region
of interest by summing in each region of interest the edge
gradients of those pixels of the respective region of interest
having the same orientation;
determining a rib edge based on the cumulative edge gradients
determined in the preceding step, comprising,
determining which of said cumulative edge gradients has a largest
sum within a predetermined range of angles; and
determining said rib edge based on a location of the region of
interest having in its respective cumulative edge gradient said
largest sum within said predetermined range of orientations.
19. A method as recited in claim 18, wherein a vertical direction
in said chest radiograph is a direction generally parallel to a
subject's spine and a horizontal direction is a direction
perpendicular to said vertical direction, and said step of defining
said pneumothorax region of interest comprises:
determining vertical profiles in regions of both lungs of said
chest radiograph;
fitting a predetermined model function to said vertical profiles to
obtain initial estimates of said rib edges in said radiograph;
and
defining first regions of interest in said radiograph which include
regions containing said initial estimates.
20. A method as recited in claim 18, wherein said step of removing
said rib edges comprises:
fitting said detected rib edges with a predetermined function
providing a rib edge curve for each of said detected rib edges;
and
removing from said pneumothorax region of interest first pixels
corresponding to said rib edge curves and a predetermined number of
second pixels surrounding each of said first pixels on each of said
rib edge curves.
21. A method as recited in claim 18, further comprising:
edge enhancing said pneumothorax region of interest by determining
edge gradients and corresponding orientations for each pixel in
said pneumothorax region of interest and assigning data values to
pixels having edge gradients at predetermined orientations and
exceeding predetermined thresholds;
defining islands in said edge-enhanced pneumothorax region of
interest;
determining a maximum magnitude of said edge gradients and a sum of
magnitudes of said edge gradients in each island in said
edge-enhanced pneumothorax region of interest;
eliminating first islands having a sum of magnitudes of said edge
gradients below a first predetermined value in said edge-enhanced
pneumothorax region of interest; and
eliminating second islands having a maximum of magnitudes of said
edge gradients above a second predetermined value in said
edge-enhanced pneumothorax region of interest.
22. A method as recited in claim 21, wherein:
said step of eliminating said first islands comprises removing
noise components from said edge-enhanced pneumothorax region of
interest; and
said step of eliminating said second islands comprises removing
remaining rib edge components from said edge-enhanced pneumothorax
region of interest.
23. A method as recited in claim 18, wherein said step of detecting
said curved line comprises:
using a Hough transform to detect said curved line corresponding to
said pneumothorax;
obtaining a strongly accumulated area having accumulated values in
a r-.theta. parameter space using said Hough transform; and
detecting said pneumothorax by applying an inverse Hough transform
to selected points in said strongly accumulated area in said
r-.theta. parameter space.
24. A method as recited in claim 23, further comprising:
eliminating in said strongly accumulated values a first portion of
said accumulated values below a first predetermined threshold;
and
eliminating in said strongly accumulated area a second portion of
said accumulated values above a second predetermined threshold.
25. A system as recited in claim 24, wherein:
said step of eliminating said first portion of said accumulated
values comprises removing noise components from said strongly
accumulated area; and
said step of eliminating said second portion of said accumulated
values comprises removing rib edge components from said strongly
accumulated area.
26. A system for detection of a pneumothorax in a chest radiograph,
comprising:
means for defining rib edge regions of interest and a pneumothorax
region of interest;
first means for detecting rib edges in said rib edge regions of
interest in said radiograph;
second means for removing said rib edges and respective adjacent
areas from said pneumothorax region of interest;
third means for detecting a curved line in said enhanced
pneumothorax region of interest having said rib edges removed;
and
fourth means for producing representation of said pneumothorax
based on the detected curve line;
wherein said first means comprises:
means for determining edge gradients and corresponding orientations
of the determined edge gradients for each of a plurality of pixels
included in each of said rib edge regions of interest;
means for determining a respective cumulative edge gradient for
each rib edge region of interest by summing the edge gradients of
pixels of the respective rib edge region of interest which have the
same orientation;
means for determining a rib edge based on said respective
cumulative edge gradients, comprising,
means for determining which of said respective cumulative edge
gradients has a greatest sum within a predetermined range of
orientations, and
means for determining said rib edge based on a location of the
region of interest having in its respective cumulative edge
gradient said largest sum in said predetermined range of
orientations.
27. A system as recited in claim 26, wherein a vertical direction
in said chest radiograph is a direction generally parallel to a
subject's spine, a horizontal direction is a direction
perpendicular to said vertical direction, and said means for
defining regions of interest comprises:
means for determining vertical profiles in both lung regions of
said chest radiograph;
means for fitting a predetermined model function to said vertical
profiles to obtain initial estimates of said rib edges in said
radiograph; and
means for defining said regions of interest in said radiograph as
regions containing said initial estimates.
28. A system as recited in claim 26, wherein said second means for
removing said rib edges comprises:
means for fitting said detected rib edges with a predetermined
function providing a rib edge curve for each of said detected rib
edges; and
means for removing from said pneumothorax region of interest first
pixels corresponding to said rib edge curves and a predetermined
number of second pixels surrounding each of said first pixels on
each of said rib edge curves.
29. A system as recited in claim 26, further comprising:
means for edge enhancing said pneumothorax region of interest by
determining edge gradients and corresponding orientations for each
pixel in said pneumothorax region of interest and assigning data
values to those pixels having edge gradients at predetermined
orientations and exceeding predetermined thresholds;
means for defining islands in said edge-enhanced pneumothorax
region of interest;
means for determining a maximum magnitude of said edge gradients
and a sum of magnitudes of said edge gradients in each island in
said edge-enhanced pneumothorax region of interest;
means for eliminating first islands having a sum of magnitudes of
said edge gradients below a first predetermined value in said
edge-enhanced pneumothorax region of interest; and
means for eliminating second islands having a maximum of magnitudes
of said edge gradients above a second predetermined value in said
edge-enhanced pneumothorax region of interest.
30. A system as recited in claim 29, wherein:
said means for eliminating said first islands comprises means for
removing noise components from said edge-enhanced pneumothorax
region of interest; and
said means for eliminating said second islands comprises means for
removing remaining rib edge components from said edge-enhanced
pneumothorax region of interest.
31. A system as recited in claim 26, wherein said third means
comprises:
means for using a Hough transform to detect said curved line
corresponding to said pneumothorax;
means for obtaining a strongly accumulated area having accumulated
values in a r-.theta. parameter space using said Hough transform;
and
means for detecting said pneumothorax by applying an inverse Hough
transform to selected points in said strongly accumulated area in
said r-.theta. parameter space.
32. A system as recited in claim 31, further comprising:
means for eliminating a first portion of said accumulated values in
said strongly accumulated area below a first predetermined
threshold; and
means for eliminating a second portion of said accumulated values
in said strongly accumulated area above a second predetermined
threshold.
33. A system as recited in claim 32, wherein:
said means for eliminating said first portion of said accumulated
values comprises means for removing noise components from said
strongly accumulated area; and
said means for eliminating said second portion of said accumulated
values comprises means for removing rib edge components from said
strongly accumulated area.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
The present invention is related to commonly-owned U.S. Pat. No.
4,851,984 and application Ser. No. 07/275,720, filed Nov. 23, 1988,
now U.S. Pat. No. 5,072,384, the disclosures of which are
incorporated herein by reference.
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to the computerized quantitative
analysis of digital chest radiographs, and in particular to a
method and system for detection of posterior ribs and pneumothorax
in digital chest radiographs.
2. Discussion of the Background
It is commonly believed that the rib structure in chest radiographs
provides a frame of reference for quantitative analysis of digital
chest images, such as the analysis of temporal changes between
successive digital chest images. For example, in the case of a lung
cancer that develops during the interval between two chest X-ray
examinations, a temporal subtraction technique has been used to
improve detectability, provided that the rib structures in the two
images may be matched by converting the locations of ribs in one
image to those of another, as suggested by Kinsey et al
"Application of Digital Image Change Detection to Diagnosis and
Follow-up of Cancer Involving Lungs," Proceedings of SPIE 70,
99-112 (1975).
It has also been shown that many false-positives occur at ribs and
rib crossings in computerized detection of lung nodules in a single
frontal chest radiograph. M. L. Giger et al "Pulmonary Nodules:
Computer-Aided Detection in Digital Chest Images," RadioGraphics
10, 41-51 (1990), and Yoshimura et al "Analysis of
Computer-Reported False-Positive Detections of Lung Nodules in
Digital Chest Radiography," Med. Phys. 17, 524 (P) (1990). For
quantitative analysis of lung textures related to interstitial
diseases, many regions-of-interest (ROIs) need to be selected
automatically in the intercostal spaces. Thus, accurate knowledge
of rib locations is essential for the development of a reliable
method for automated selection of ROIs.
In the prior art, various methods have been developed for automated
rib detection. Generally, these methods have attempted to detect
local rib edges while applying some anatomic knowledge to construct
the rib structure. However, these methods are still far from being
ready for practical use on clinical chest images, and some
difficulties in automated rib detection remain. For example, chest
images contain radiographic noise and also many confusing edges due
to blood vessels, bronchi, lung texture, lesions and artifacts. In
addition, rib contrast is commonly low, and rib edges are often
ill-defined because of poor image quality. Accordingly, the
signal-to-noise ratio of rib structures is generally low.
Similarly, accurate knowledge of rib locations is useful for
development of automated techniques for the detection of
pneumothorax. Pneumothorax is a condition caused by an accumulation
of air or gas in the pleural cavity, which occurs as a result of
disease or injury. Radiographic detection of pneumothorax is
commonly based on a subtle, fine curved-line pattern in the apical
lung region, a dark pleural air space against the chest wall due to
increased transparency, and a lack of lung structure between the
rib cage and the pneumothorax pattern. Although pneumothoraces are
clinically important abnormalities, it is difficult to detect them,
in part because there is overlap between the pneumothoraces and the
ribs and clavicle. Prior art techniques have been used to enhance
the pneumothorax pattern by use of digital processing of chest
images. However, no attempt has been made to detect pneumothorax
automatically by means of a computer. Computerized, automated
detection of subtle pneumothorax patterns would be helpful for the
diagnosis made by radiologists in that they will be alerted to a
potential subtle lesion.
SUMMARY OF THE INVENTION
Accordingly, one object of this present invention is to provide a
novel method and system using edge gradients and their orientations
determined from a digital chest radiograph to automatically detect
edge structure of anatomic features in the radiograph.
Another object of the present invention is to employ statistical
analysis of edge gradients and their orientations in a digital
chest radiograph for automatic delineation of posterior rib edges
and the subtle curved-line pattern of pneumothorax in the apical
lung region.
These and other objects of the present invention are obtained by a
method and system for detecting anatomic features in a digital
chest radiograph automatically, in which ROIs are defined in the
digital chest radiograph in areas containing edge structure desired
to be detected and edge gradients and their corresponding
orientations are obtained using a Sobel operator for a plurality of
pixels included in each ROI. The maximum of the edge gradients in
each ROI is determined and used to produce a representation of the
anatomic feature. Typically, according to the invention, the
maximum edge gradients are summed to form cumulative maximum edge
gradients which are plotted as a function of their orientation to
obtain a histogram. The histogram provides an easily discernable
and accurate detection of the desired edge of the anatomic
feature.
In a first embodiment according to the present invention, the
method and system for automatic delineation of anatomic edge
structure are used to detect posterior rib edges. Vertical profiles
are determined in each lung in the digital chest radiograph and a
predetermined model function is fit to the vertical profiles to
obtain initial estimates of the positions of the rib edges in the
chest radiograph. ROIs are defined in the vicinity of the estimated
rib edges and edge gradients and their corresponding orientations
are determined using a Sobel operator. A maximum edge gradient for
each ROI is determined which corresponds to a detected rib edge.
Cumulative maximum edge gradients are determined and plotted as a
function of their orientation to obtain a histogram. The resulting
histogram contains peaks which correspond to detected rib edges.
The detected rib edges are also fit with an elliptical function
providing a representation of the entire rib structure.
In some cases, due to low signal-to-noise ratio and interference
with other anatomic features, all rib edges are not detected. These
missing rib edges may be estimated by analyzing the elliptical
functions of detecting rib edges and using anatomic knowledge of
the variation of rib edges with position within a chest radiograph.
The missing rib edges may also be detected by estimating their
position from the previously detected rib edges and then performing
again the gradient-histogram analysis in the vicinity of the
estimated rib edge.
In a second embodiment according to the present invention,
automatic delineation using edge gradients and their orientations
are implemented to detect pneumothorax in the apical lung region.
After the detection of rib edges in a chest radiograph, ROIs are
defined in the apical region of each lung in the radiograph. Edge
gradients and their corresponding orientations are determined for a
plurality of pixels included in each of the ROIs. The ROIs are
enhanced using the edge gradients and the corresponding
orientations to accentuate edge structure in the ROIs. The detected
rib edges are removed from the edge-enhanced ROIs by removing the
pixels corresponding to a curved line representing the detected rib
edge. A number of pixels surrounding each pixel on a curved line
representing a detected rib edge are also removed. Curved-line
structure remaining in the edge-enhanced ROIs having rib edges
removed are detected using a Hough transform. A strongly
accumulated area having accumulated values in a r-.theta. parameter
space is obtained. A pneumothorax pattern is detected using the
inverse Hough transform on selected points in the strongly
accumulated area in the r-.theta. parameter space.
The detection of the pneumothorax can be enhanced by eliminating
accumulated values in the strongly accumulated area below a first
predetermined threshold which removes noise components from the
strongly accumulated area. Secondly, accumulated values above a
second predetermined threshold are removed which removes components
due to rib edges in the strongly accumulated area. The
signal-to-noise ratio of the detected pneumothorax pattern is
improved.
BRIEF DESCRIPTION OF THE DRAWINGS
A more complete appreciation of the present invention and many of
the attendant advantages thereof will be readily obtained as the
same becomes better understood by reference to the following
detailed description when considered in connection with the
accompanying drawings, wherein:
FIG. 1(a) is an illustration of manual fitting of peripheral rib
edges using elliptical functions in a digital chest image;
FIGS. 1(b) and 1(c) are graphical representations of the dependence
of fitted parameters of the elliptical functions on the location of
ribs;
FIG. 2 is an illustration of vertical profiles in a chest
radiograph;
FIG. 3 is an illustration of a chest image with selected ROIs in
the right lung;
FIG. 4 is a graphical representation of the distribution of edge
gradients and their orientations at all 32 pixels included in each
of the five ROIs shown in FIG. 3;
FIG. 5(a) is a graphical representation of the relationship between
the edge gradients and their orientations for the five ROIs shown
in FIG. 3;
FIG. 5(b) is a histogram of the cumulative gradient as a function
of orientation;
FIG. 6(a) is an illustration of a chest radiograph with rib edges
obtained with a Sobel filter;
FIG. 6(b) is an illustration of a chest radiograph with rib edges
obtained through gradient-histogram analysis;
FIG. 7 is a graphical illustration of the determination of a lower
edge of a posterior rib;
FIG. 8 is an illustration of a chest radiograph comparing the
detection of rib edges from vertical profiles (dark markers) and
rib edges detected by gradient-histogram analysis (light
markers);
FIGS. 9(a)-(c) are illustrations of a chest radiograph with
automated delineation of posterior ribs superimposed on the chest
image exemplifying in (a) and (b) "good" cases and in (c) an
"acceptable" case;
FIG. 10 is an illustration of a chest image with pneumothorax in
the apical lung, a ROI marked with a white rectangle selected over
a fine line pattern due to pneumothorax;
FIG. 11 is a graphical representation of the average image profile
across the pneumothorax pattern of FIG. 10 indicating sharp and low
edge contrast;
FIG. 12 is an illustration of a chest radiograph with two ROIs
selected in apical lungs;
FIG. 13(a) is an illustration of the two ROIs selected in apical
lungs of FIG. 12 after edge enhancement;
FIG. 13(b) is an illustration of the edge enhanced image of FIG.
13(a) after removal of posterior rib edges;
FIG. 13(c) is an illustration of the image shown in FIG. 13(b)
after removal of image noise components;
FIG. 14 is an illustration of a chest radiograph depicting
automated computerized detected posterior rib edges;
FIG. 15 is a graphical representation illustrating the relationship
between the maximum edge gradient to the total edge gradient of an
island corresponding to an isolated dark area in FIG. 13(b);
FIGS. 16(a) and (b) are graphical representations illustrating two
straight lines a and b in an x-y plane which are represented by two
points (r.sub.a, .theta..sub.a) and (r.sub.b, .theta..sub.b),
respectively, in the Hough transform parameter space, and the
points m and n in the x-y plane are represented by two sinusoidal
curves in the parameter space;
FIG. 17 is a graphical representation of a curved-line simulation
of a pneumothorax pattern, each of five points marked by an X being
connecting points of two lines obtained by the inverse Hough
transform of two adjacent points in a strongly accumulated
elongated area of the parameter space shown in FIG. 18;
FIG. 18 is a representation of the Hough transform of the curved
line simulating a pneumothorax pattern shown in FIG. 17, with the
strongly accumulated elongated area represented by six points each
marked by a plus symbol;
FIG. 19 is a representation of the Hough transform of the edge
enhanced image in FIG. 13(c);
FIG. 20 is a graphical representation of the distribution of
accumulated values for islands due to rib edges, pneumothorax and
image noise;
FIG. 21 is a graphical representation of the distribution of total
accumulated noise in outer lung space due to pneumothorax, rib edge
and image noise;
FIG. 22 is an illustration of a chest radiograph with detected
points of a pneumothorax pattern depicted on the chest image
obtained from the inverse Hough transform of the five points
representing a strongly accumulated elongated area in the Hough
transform parameter space shown in FIG. 19;
FIGS. 23(a) and (b) are illustrations of chest images with correct
automated detections of pneumothorax in apical lung;
FIG. 24 is a block diagram of a first embodiment of the system
according to the present invention;
FIG. 25 is a block diagram of the anatomic feature detector shown
in FIG. 24;
FIG. 26 is a block diagram of a second embodiment of the system
according to the present invention for detection of rib edges;
and
FIG. 27 is a block diagram of a third embodiment of the system
according to the present invention for detection of
pneumothorax.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
Referring now to the drawings, and more particularly to FIG. 1(a)
thereof, wherein a comparison of elliptical functions and actual
rib edges in a chest image is illustrated, a first embodiment of
the present invention will be described. The chest images
illustrated in the drawings are representative of fifty frontal
chest images selected mainly from routine cases by the inventors.
The digital images were obtained by digitizing the chest
radiographs with a Konica laser scanner using a pixel size of 0.175
mm.times.0.175 mm and a matrix size of 2,000.times.2,430. Digital
images with a small matrix (1,000.times.1,215) and a large pixel
size (0.35 mm.times.0.35 mm) were also used for computational
simplicity. The gray scale used was 10 bits. Further, only the
posterior ribs in the peripheral portions of the lungs were
considered because accurate delineation of peripheral posterior
ribs is adequate for utilizing rib locations in quantitative
analysis of digital chest images.
FIG. 1(a) illustrates the comparison of elliptical functions and
actual rib edges in a chest image. The location of the peripheral
posterior ribs in chest radiographs is usually very similar in
shape to a portion of an ellipse, implying that an elliptical
function may be used for fitting of the detected edges of
peripheral ribs and the smoothing of detected data points. To that
end, each elliptical function was generated to match rib edges
visually by manual selection of three parameters, namely, the two
diameters and the location of the center of the ellipse.
Some parameters of the fitted ellipse are closely related to the
geometry of the chest image. For example, the center of the ellipse
is at the vertical midline of the chest image, and one-half of the
long diameter corresponds to the distance between the rib cage edge
and the midline. FIGS. 1(b) and 1(c) show the dependence of the
fitted parameters on the location of the ribs. Implied from the
geometry of the chest image, the parameters change gradually from
one rib to the next. The parameters for the right lung are
generally similar to those for the left lung. In addition, the
parameters for the upper edges of the ribs (open symbols) are
closely related to those for the lower edges of the ribs (closed
symbols).
According to the first embodiment of the present invention,
vertical profiles in the lung regions are determined in order to
obtain the rib cage edges by analysis of the horizontal signature,
which is then fitted to a third order polynomial function. As shown
in FIG. 2, ten profiles are selected which extend in the same
direction as the rib cage edges in both lung regions. The profiles
are located between approximately 95% and 50% of the distance from
the midline of the chest to the corresponding fitted smooth curves
defining the rib cage edge boundaries.
Each vertical profile is then fitted with a shift-variant
sinusoidal function, which is given by the equation
where f(x) corresponds to the background trend corrected profile, A
is the amplitude of the vertical profile, .phi. is a phase term,
and u(x) is the spatial frequency at position x. This spatial
frequency corresponds to the reciprocal of the rib plus intercostal
distance. The rib plus intercostal distance is assumed to change
linearly with the position x. The fitting using the shift-variant
sinusoidal function obtains initial estimates of the locations of
the upper and lower edges of the posterior ribs. This initial
estimation of the rib edge locations is performed in the same
manner as that disclosed in Doi et al (U.S. Pat. No.
4,851,984).
From these initial estimates of the locations of the upper and
lower edges of the posterior ribs, subtle continuous rib edges are
detected using a technique termed gradient-histogram analysis (GHA)
wherein statistical analysis of edge gradients and their
orientations of all pixels in small selected ROIs is performed. GHA
makes it possible to determine whether a rib edge is included in a
particular ROI and, if so, whether the rib edge is located in the
upper or lower margin of the posterior or anterior rib.
FIG. 3 shows five ROIs in the right lung, which were selected along
the upper and lower margins of both posterior and anterior ribs,
and also from the intercostal space. Each ROI is 8.times.4 pixels
and rectangular in shape (longer in the horizontal direction) as
most rib edges commonly lie in the horizontal direction. However,
other sizes and shapes of ROIs may be used to obtain similar
results. The selection of an appropriate size and shape of the ROI
should be based on the parameters which yield the best results.
The distributions of gradients and their orientations at all pixels
included in the five ROIs of FIG. 3 are shown in FIG. 4. The
gradient and its orientation denote the maximum edge gradient and
its orientation at each pixel, which are obtained from two edge
gradients in both vertical and horizontal directions using a Sobel
operator or other edge operators may also be used. A Sobel operator
is described in Gonzalez et al, "Digital Image Processing", 2nd Ed.
(Addison-Wesley, Boston, Mass. 1987), pp. 176-179, which is herein
incorporated by reference. The determination of the edge gradients
is not limited to the use of a Sobel operator. Other edge operators
may be used, such as first and second derivatives, Laplacian
operator etc.
The gradients and their orientations are plotted in a polar
coordinate system and the magnitude of the maximum edge gradient
corresponds to the distance from the origin to a point marked in
the diagram. The orientation of the edge gradient is represented by
the angle from the horizontal axis on the positive x axis. These
distributions may also be obtained by plotting the edge gradients
in the horizontal and vertical direction on the abscissa and
ordinate, respectively. It is apparent in FIG. 4 that four
different rib edges are clearly distinguished from each other and
are represented by points in four different quadrants. The
intercostal space can also be distinguished from the rib edges,
because the intercostal space edge gradients are much smaller than
those of the rib edges.
A background trend correction is useful in separating edge
gradients due to rib edges and those in the intercostal space, when
a non-uniform background trend due to chest wall and gross lung
anatomy is involved in each ROI, which is commonly the case. The
background trend is estimated by fitting pixel values in each ROI
using a two-dimensional polynomial function. The fitted function is
subtracted from pixel values in the ROI to provide the
trend-corrected ROI. Edge gradients and their orientations are then
determined using the Sobel operator. This technique tends to
decrease the magnitude and scatter of edge gradients obtained from
the intercostal space, and thus is effective in improving the
correct identification of rib edges. This background trend
correction technique is similar to that described in Doi et al
(U.S. Pat. No. 4,851,984).
In FIG. 5(a), the data in FIG. 4 are displayed with the gradient
and orientation plotted on the abscissa and ordinate, respectively.
FIG. 5(b) shows the distributions (or histograms) of cumulative
gradients as a function of orientation. Cumulative gradients are
formed by summing of the edge gradients at each orientation. With
the use of the cumulative gradients, it is possible to efficiently
identify the presence or absence of a rib edge as well as the
nature of the rib edge.
FIG. 6(a) illustrates a chest image with many edges, oriented in
various directions obtained with a Sobel operator. Not only rib
edges but also many unwanted edges are apparent when edge gradients
in all directions are considered. This indicates that the
corresponding signal-to-noise ratio of the rib edges is not very
high. FIG. 6(b) shows an image of edges due to the lower margin of
posterior ribs obtained with GHA. With GHA, the signal-to-noise
ratio of the rib edges is improved, and the lower margins of the
posterior ribs are well defined. Although GHA can be applied to the
entire chest image, it would be time consuming for .computation.
Therefore, it is desirable to use GHA only on relatively small
regions adjacent to the initially estimated rib edges.
An example of the use of GHA in the detection of rib edges is shown
in FIGS. 7 and 8. First, the chest image is segmented by continuous
rectangular ROIs in the vicinity of the initial edges which are
obtained by fitting of the vertical profiles with a shift-variant
sinusoidal function, which was described previously. Nineteen ROIs
are placed along the vertical direction centered at the initially
estimated rib edge. The size of the ROI is desirably 8.times.4
pixels, but other sizes are possible, and each ROI overlaps
one-half of the adjacent ROI. GHA is then performed on all ROIs. A
reliable rib edge is finally detected by finding the ROI which
provides the largest cumulative edge gradient within a given range
of edge orientations.
FIG. 7 illustrates the determination of the lower edge of a
posterior rib. The actual rib edge was detected at the upper side
of the initially estimated edge. FIG. 8 shows the comparison of rib
edges as detected initially from the vertical profiles and accurate
rib edges as detected by GHA, and are indicated by dark and light
markers, respectively. Almost all edges in the mid and upper ribs
are detected accurately by GHA. However, some rib edges in the
lower lung region are not detected, because these edges are
obscured due to overlap with pulmonary vessels.
Although the majority of rib edges are detected correctly by GHA as
described above, some incorrect edge data may be generated and also
some edges may not be found due to the extremely low
signal-to-noise ratios of some local rib edges. Therefore, in order
to obtain estimates of smooth rib edges for those having low
signal-to-noise ratios, two methods are employed, namely,
horizontal alignment of detected edges and fitting of the detected
edges with an elliptical function. Horizontal alignment groups
detected edge points into different rib edges. This is accomplished
by taking into account the orientation of its gradients for
distinction of upper and lower rib edges, and also by examining the
vertical position of the detected edge as a function of the
horizontal distance. The vertical position of the rib edge
decreases gradually towards the rib cage edge in the periphery of
the lungs.
When the detected edge points have been grouped together, the rib
edges in each group are fitted with an elliptical function, thus
producing sets of fitting parameters for a number of rib edges from
the top to the bottom rib. Since these parameters change gradually
depending upon the location of the rib, as shown in FIG. 1(b), it
is possible to estimate a set of fitting parameters for a rib edge
which has not been detected. If necessary, the missing rib edges
are searched for again by GHA near the potential rib locations
estimated from the set of fitting parameters obtained above. This
fitting is effective to correct some errors that may have occurred
in the initial GHA rib detection.
As a second example of this embodiment, GHA including the
horizontal alignment of detected edges and fitting with an
elliptical function was performed on 50 chest radiographs. The CPU
time using a DEC VAX 3500 computer required for each GHA analysis
is approximately 60 seconds for each case. The detected edges are
compared with actual rib edges and are classified into three
categories, "good", "acceptable" and "inadequate", based on the
goodness of fit (or delineation) of detected rib edges compared
with actual rib edges.
These categories are defined based upon the fraction of the
correctly delineated rib edges relative to the number of desired
rib edges in a chest image. The number of desired rib edges which
are located in the lung regions and are considered to be required
for the acquisition of rib structure, is usually from 20 to 24 (5
to 6 ribs above the diaphragm in each hemithorax). Thus, the good
category corresponds to the correct delineation of more than 80% of
the total number of desired rib edges, an acceptable delineation is
60% to 80%, and inadequate is the delineation of less than 60% of
the desired rib edges.
FIGS. 9(a) and 9(b) show two examples of a good case. The results
of automated delineation of posterior ribs are marked by white
curves, which are superimposed on the chest image. The annotated
rib edges are generally in good agreement with actual rib edges. In
FIG. 9(a), the edges of the fourth to ninth rib in both lungs are
determined correctly except for the lower edge of the ninth rib in
the left lung. This was caused by erroneous initial estimation of
the rib edges and the subsequent detection of blood vessel edges.
In FIG. 9(b), although the image contrast is relatively low, five
ribs in both lungs, except for the upper edge of the fourth rib,
are correctly determined. However, some lower ribs with very
obscure edges are not detected because of incorrect initial
estimations for these ribs. In contrast, FIG. 9(c) shows an example
of an acceptable case. Here, the edges of the fourth to sixth ribs
in the left lung are confused because of some errors in the initial
estimation of the rib edges and in the horizontal alignment of
detected incorrect edges. Also, because of other incorrect
horizontal alignment, one false-positive result occurred in the
upper right lung.
The overall results for the 50 cases indicated 37 (74%) good cases,
8 (16%) acceptable cases, and 5 (10%) inadequate cases. The number
of false-positive delineations were, on the average, 1.2, 2.5 and
1.2 per chest image for good, acceptable, and inadequate cases,
respectively. These false-positives usually occur near the
correctly detected edges, and some are located on or below the
diaphragm edge.
From these results on the 50 chest radiographs, it is apparent that
the correct delineation of rib edges is strongly dependent on the
initial estimation of rib edges, and that a false-positive can
occur due to incorrect horizontal alignment of edge data. For the
lower edges of ribs in the lower lung, when an error in the initial
estimation occurs, GHA tends to detect edges of blood vessels, and
thus the rib edge detected tends to be shifted below the actual rib
edge near the mediastinum.
In some cases, incorrect fitting of rib edges with the elliptical
function occurred as a result of erroneous determination of rib
cage edges caused by breast shadows over the lower lung. However,
the fitting of edge data with an elliptical function is effective
not only in yielding good agreement between a detected rib and
actual rib edges, but also for estimation of obscure rib edges,
which cannot be easily detected, when the dependence of fitted
parameters on the elliptical function is utilized. The missing rib
edge can be estimated based on the parameters of the elliptical
functions obtained for the other edge of the same rib. The missing
rib may also be estimated with parameters of elliptical functions
derived from other ribs. These two estimation techniques provide
generally satisfactorily results. However, false-positives are
produced occasionally when the order of aligned edge data does not
correctly match the order of the actual rib.
Most false-positives may be eliminated by modification of the first
embodiment of the present invention. Estimation of undetected rib
edges based upon the dependence of the fitting parameters may be
improved by taking into account other radiographic knowledge of
ribs such as widths of the ribs and intercostal spaces.
In a second embodiment according to the present invention, GHA is
used to detect pneumothorax in the apical lung. An example of
pneumothorax in a postero-anterior (PA) chest radiograph is shown
in FIG. 10, and the average image profile of the pneumothorax
pattern in the ROI is shown in FIG. 11. The average image profile
of FIG. 11 indicates the distribution of pixel values in a
direction perpendicular to the pneumothorax pattern. Since many
profiles are averaged over the width of approximately 50 pixels
(length of the short side of the ROI), radiographic noise is almost
eliminated in the average profile. However, if the distribution
along a profile one pixel wide is illustrated, it becomes
impossible to recognize an edge due to the pneumothorax pattern
because of the high level of radiographic noise.
It is apparent in FIG. 10 that the pneumothorax pattern contains a
subtle sharp edge with very low contrast, and the edge of the
pneumothorax has a slightly larger gradient than that of the
background trend. The gradient of the edge of pneumothorax is in
the direction opposite to that of the background trend. This
corresponds to the fact that the average density in the apical lung
field decreases gradually toward the upper outer corner of the rib
cage edge, but increases sharply when crossing the pneumothorax
pattern. Hence, the pixel value is inversely proportional to the
optical density.
Eight chest images with similar pneumothorax patterns were analyzed
and the results are summarized in Table 1.
TABLE 1 ______________________________________ Summary of Measured
Edge Width and Contrast of Pneumothorax Patterns in Eight PA Chest
Radiographs Edge width Edge contrast Case No. (mm) (pixel units)
______________________________________ 1 0.30 14.9 2 0.41 16.2 3
0.24 8.4 4 0.46 15.5 5 0.48 16.1 6 0.22 10.5 7 0.40 11.3 8 0.24
14.5 average 0.34 .+-. 0.10 13.4 .+-. 2.9 0.040 .+-. 0.009 (optical
density) ______________________________________
The average edge width and the average contrast of the sharp line
structures due to pneumothoraces are 0.34 mm and 0.04 in optical
density units, respectively. The edge width obtained from the image
profile as shown in FIG. 11 is only an estimate which tends to be
an over-estimate because of the averaging of many profiles across
the pneumothorax pattern and also due to a relatively large pixel
size (0.175 mm). It is therefore likely that an actual edge widths
of pneumothorax patterns are much smaller (and thus very sharp)
than the results of the measurements shown in Table 1.
In the second embodiment according to the present invention, the
rib cage edge, the top and bottom of the lung, which are useful
parameters indicating the lung regions, can be obtained using the
analysis of the horizontal signature in a manner such as that
disclosed by Doi et al (U.S. Pat. No. 4,851,984). In order to
facilitate an efficient detection of the pneumothorax pattern, ROIs
are determined along the rib cage edge in both apical areas of the
lungs where subtle pneumothoraces commonly appear.
An example of such ROIs is shown in FIG. 12. The right and left
vertical boundaries of the ROI are selected at, for example, 95%
and 60%, respectively, of the distance from the midline of the
chest image to the corresponding fitted smooth curve defining the
boundary of the rib cage edge. The upper and lower horizontal
boundaries of the ROI are determined at, for example, 10% and 40%,
respectively, of the distance from the detected top to the bottom
of the lung. The criteria for selecting the ROI boundaries are
determined empirically. The larger the area of the ROI, the more
difficult the detection of the pneumothorax pattern because of an
increase in the number of rib structures in the ROI. However, the
selection of a very small ROI may fail to include a sufficiently
large portion of the pneumothorax pattern, making its detection
more difficult.
The subtle pneumothorax pattern in the ROI is then enhanced. First,
the orientation of the edge gradient is determined by the use of a
Sobel filter in two orthogonal directions. To obtain accurate edge
gradients of the pneumothorax pattern, the kernel size of the Sobel
filter was chosen to be 3.times.3 pixels by taking into account the
edge width of the pneumothorax patterns. FIG. 13(a) illustrates an
edge-enhanced image based on the edge gradient within a certain
limited range of its orientation. The range is selected such that
the optical density in the chest image increases towards the upper
outer corner in the apical lung. This density change includes edges
due to the pneumothorax pattern. The pneumothorax is clearly
visualized as a thin dark dotted line in FIG. 13(a). However, some
rib edges are also enhanced and result in relatively thick dark
lines being present in the enhanced image, since the orientations
of the edge gradients for some rib edges are in the same range as
those for the pneumothorax pattern. Also, many isolated noise
components are due to radiographic mottle included in the chest
image. Thick white lines as seen by lack of noise components are
due to anterior ribs and clavicles, the orientations of their edge
gradients being in a direction roughly perpendicular to those of
the pneumothorax pattern.
The rib edges are then removed from the enhanced image. In order to
remove these edges, the computer output from the analysis of the
automated delineation of the posterior ribs (as shown in FIG. 14)
is used, which is described above and omitted here for brevity. To
distinguish a pneumothorax pattern from rib edges, the rib edges
and adjacent areas within a certain width, for example .+-.20
pixels around the fitted elliptical curves, were removed. This
width is determined empirically, and the result is shown in FIG.
13(b). It is apparent from FIG. 13(b) that most rib edges which are
seen as thick dark lines in FIG. 13(a) are now replaced by white
bands. However, there are still many noise components and some
residual rib components remaining in the image. It is desirable to
remove these components, which is done by thresholding. The
threshold level used in the removal of these components is
determined by analysis of each component, which will be referred to
as an island, that corresponds to a set of eight neighboring
connected pixels.
FIG. 15 illustrates the relationship between the maximum and total
edge gradient for islands included in edge enhanced images. The
maximum edge gradient corresponds to the largest edge gradient
value within the island. The total edge gradient corresponds to the
total (or summation) of all edge gradient values over the entire
area of the island. The maximum edge gradients of the residual rib
edges are usually greater than those of the pneumothorax, as shown
by the circles and triangles in FIG. 15, as rib edges have
generally greater contrast than do pneumothorax edges.
Also shown in FIG. 15 is the total edge gradients of image noise
components, which are generally smaller than those of either rib
edges or pneumothoraces. This is because the area of the island due
to an image noise component is generally smaller than that due to a
pneumothorax. Therefore, to distinguish pneumothorax from the rib
edges and the image noise components, all islands except those in
the lower right rectangular area indicated by the dashed lines in
FIG. 15 are eliminated. FIG. 13(c) shows the edge-enhanced image
after removal of the rib edges and image noise components. The
pneumothorax pattern is recognizable and the signal-to-noise ratio
of the pneumothorax pattern is improved.
Points located on a curved line in the image which is suspected of
being a pneumothorax are detected using a Hough transform. The
Hough transform is described by Gonzales et al, ibid, and omitted
here for brevity. Using the Hough transform, straight lines in an
x-y plane can be represented by points in an r-.theta. parameter
space as illustrated in FIGS. 16(a) and 16(b). The distance r
equals the perpendicular distance from the origin to the straight
line and the angle .theta. is the angle from the x-axis to the
perpendicular line to this straight line. The relationship between
the coordinates in the two systems is therefore described by
Using this relationship, a point in the x-y plane becomes a
sinusoidal curve in the parameter space. Therefore, when there are
many (but a finite number of) discrete points on the same straight
line in an x-y plane, the line can be determined or estimated from
analysis of accumulated values in the parameter space. All
sinusoidal curves due to all discrete points go through a point in
the parameter space and this point represents a straight line in
the x-y plane. The conversion of a point or points in the parameter
space to a line or lines in the x-y domain may be called the
inverse Hough transform.
The Hough transform is used to detect the pneumothorax pattern as
described in the following manner. A curved line simulating a
pneumothorax pattern is shown by a thick curve in FIG. 17, and its
corresponding Hough transform is shown in FIG. 18. The points lying
on the curved line in FIG. 17 are represented by many sinusoidal
curves and thus produce a strongly accumulated elongated (dark)
area in the parameter space as shown in FIG. 18. If this
accumulated area is approximated by a single point with a very
large accumulated value, then the curved line should have been
nearly a straight line. This is an obvious consequence of the basic
relationship between the paired patterns in the spatial domain and
in the parameter space.
However, if the accumulated area is approximated by several
adjacent points such as those marked by a plus in FIG. 18, then the
curved line could be represented by a number of segments of
straight lines, as illustrated by the thin line segments connected
via x's in FIG. 17. Each end of the line segment marked by an x
corresponds to the connecting point of two straight lines which are
obtained from the inverse Hough transform of two adjacent points in
the parameter space in FIG. 18. In order to obtain points
representing the strongly accumulated area in the parameter space,
the distance r is determined at each angle .theta. with an
appropriate increment, yielding the maximum accumulated value above
a certain threshold value. Inverse Hough transforms of these data
points on r and .theta. provide the line segments representing a
curved line. In FIG. 18, the original thick curve is similar to the
thin connected line segments. Thus, the curved line which
represents a pneumothorax pattern can be detected using the Hough
transform.
For the detection of an actual pneumothorax pattern in a chest
image, the Hough transform is applied on the edge-enhanced image
shown in FIG. 13(c), and the inverse Hough transform is
subsequently applied to data points in the strongly accumulated
elongated area in the parameter space. In some clinical cases
however, some noise components in the edge-enhanced image produce
strongly accumulated values in the parameter space due to many
overlaps of sinusoidal curves, as illustrated in FIG. 19.
Therefore, in order to correctly identify the island (strongly
accumulated elongated area) due to the pneumothorax pattern, the
nature of the accumulated values in the parameter space are
analyzed. FIG. 20 shows the distribution of accumulated values for
islands in the Hough transform parameter space and their
corresponding locations on the .theta.-axis, obtained from 5 chest
images. The accumulated values of the residual rib components,
marked with a triangle, are generally greater than those of the
pneumothorax, which are marked by a circle. Also, the image noise
components marked by a plus tend to be smaller than the components
of the pneumothorax pattern. Therefore, the island derived from the
pneumothorax can be distinguished from islands derived from ribs
and islands derived from image noise components by thresholding at
a high and a low accumulated value, illustrated by the two dotted
lines in FIG. 20.
The total accumulated noise level in the outer lung space for a
pneumothorax pattern is zero, or nearly zero, but the noise level
for image noise and rib edges tends to have a detectable value, as
shown in FIG. 21. The Hough transform distance as shown by the
ordinate corresponds to r in the parameter space, indicating the
location of the island due to pneumothorax, rib edge or image
noise. The total accumulated noise level equals the summation of
accumulated values in the distance (at a given .theta.) range
greater than the Hough transform distance of the island defined
above. The variation in noise levels may be explained by the fact
that there is no lung structure such as small vessels and lung
texture in an outer space of the lungs associated with a
pneumothorax pattern. Therefore, when a potential pneumothorax
island is detected, the noise levels in the outer lung space are
examined to distinguish a pneumothorax pattern from other
components. As one of the most important radiographic findings of
pneumothorax by a radiologist is the lack of lung structure in the
outer space, the present invention provides extremely valuable
information to the radiologist in the diagnosis of a
pneumothorax.
An example of the detection of pneumothorax according to the second
embodiment of the invention is shown in FIGS. 22 and 23. The
analysis of these radiographs was performed using a DEC VAX 3500
computer and approximately two minutes of CPU time was required for
each case. FIG. 22 shows four detected points of a pneumothorax
pattern superimposed on the chest image, as indicated by the white
x's. The pneumothorax pattern is correctly detected. These detected
points are fitted with a second order polynomial function, and
annotated with arrows with appropriate positions as shown in FIGS.
23(a) and FIGS. 23(b). Thus, pneumothoraces can be accurately
detected by selection of edge gradients oriented only toward the
upper peripheral region of the lungs. Subtle patterns of
pneumothoraces can be identified using the Hough transform after
eliminating rib edges and image noise components. The curved lines
obtained by the inverse Hough transforms are in good agreement with
actual pneumothoraces. It is therefore possible to detect a subtle
pneumothorax by a computerized automated analysis of digital chest
radiographs which provides important information to radiologists
and improves the diagnosis of pneumothorax.
The system according to the present invention is shown in FIG. 24.
An x-ray apparatus 10 is used to acquire a chest radiograph from a
subject. The acquired radiographs are digitized using a laser
scanner, such as the previously mentioned Konica laser scanner. The
output of the laser scanner 11 is used by an ROI definer 12 to
define ROIs which include regions in the chest radiograph having
edge structure of an anatomic feature to be detected. Using the
defined ROIs, an edge gradient and orientation of the edge gradient
determiner 13 determines the edge gradient and its orientation for
each of a plurality of pixels included in the each ROI. A maximum
edge gradient detector 14 determines the maximum edge gradient in
each ROI. Based upon the maximum edge gradients detected by the
detector 14, the anatomic feature detector 15 detects the edge
structure of a desired anatomic feature, and produces a
representation thereof. The representation is then displayed on
display 17, such as a CRT.
In a second embodiment of the invention according to the present
invention, the anatomic feature detector 16 is constructed as shown
in FIG. 25. The edge gradients are accumulated by edge gradient
accumulator 21 and the histogram generator 22 forms histograms by
plotting the maximum cumulative edge gradients as a function of
their orientation. The cumulative maximum edge gradient is detected
by cumulative maximum edge gradient detector 20. Using the
cumulative maximum edge gradient data, anatomic feature edges are
detected by anatomic feature edge detector 15. The output of the
anatomic feature edge detector 15 is used by the anatomic feature
representation producer 16 to produce a representation of the
anatomic feature desired to be detected.
A second embodiment of the system according to the present
invention for the detection of ribs is shown in FIG. 26. This
system additionally includes a vertical profiler 30 for defining
vertical profiles in each lung in the chest radiograph, and a rib
edge estimator 31 which fits a shift-variant sinusoidal function to
the vertical profiles to obtain estimates of the rib edges. The
remainder of the device functions similarly to the systems shown in
FIGS. 24 and 25, with the anatomic feature edge detector 15 and the
anatomic feature representation producer 15 respectively serving as
a rib edge detector 32 and a rib edge representation producer 33.
The embodiment may also include a missing rib estimator 34 for
estimating ribs which are not detected initially. When using a
missing rib edge detector 34, the rib edge representation producer
33 produces elliptical functions fitted to each rib edge. Thus, the
missing rib edges may be estimated directly from the elliptical
functions, or edge gradients may be redetermined in areas where a
missing rib is estimated to be located by the missing rib edge
estimator 34.
FIG. 27 illustrates a further embodiment of the system according to
the present invention used to detect pneumothorax. This embodiment
includes a rib edge detector 40 which may include the structure
shown in FIG. 25. Further, after edge gradients have been detected,
the edge enhancer 41 enhances the ROI using these edge gradients.
The rib edge remover 42 removes rib edges detected by rib edge
detector 40 from the edge-enhanced ROI. A Hough transformer 43 uses
a Hough transform to detect curved line structure in the
edge-enhanced ROI having the rib edges removed. The inverse Hough
transformer 44 applied to selected points in a strongly accumulated
area formed by the Hough transform detects a pneumothorax pattern.
The pneumothorax representation producer 45 produces a
representation of the detected pneumothorax pattern which is
displayed on display 17. This embodiment may also include
thresholders 46 for removing image noise and components due to rib
edges from the edge-enhanced ROI and the Hough transform parameter
space.
The embodiments of the system shown in FIGS. 24-27 are preferably
implemented in software and controlled via a computer, but it is
also possible to implement these embodiments in a hardware
system.
Obviously, numerous modifications and variations of the present
invention are possible in light of the above teachings. It is
therefore to be understood that within the scope of the appended
claims the invention may be practiced otherwise than as
specifically described herein.
* * * * *